CN107979554B - Radio signal Modulation Identification method based on multiple dimensioned convolutional neural networks - Google Patents
Radio signal Modulation Identification method based on multiple dimensioned convolutional neural networks Download PDFInfo
- Publication number
- CN107979554B CN107979554B CN201711144077.6A CN201711144077A CN107979554B CN 107979554 B CN107979554 B CN 107979554B CN 201711144077 A CN201711144077 A CN 201711144077A CN 107979554 B CN107979554 B CN 107979554B
- Authority
- CN
- China
- Prior art keywords
- layer
- multiple dimensioned
- neural networks
- convolutional neural
- signal
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L27/00—Modulated-carrier systems
- H04L27/0012—Modulated-carrier systems arrangements for identifying the type of modulation
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L27/00—Modulated-carrier systems
- H04L27/18—Phase-modulated carrier systems, i.e. using phase-shift keying
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The present invention discloses a kind of radio signal Modulation Recognition based on multiple dimensioned convolutional neural networks, implementation step are as follows: (1) generates treated radio modulation signal；(2) two-dimentional time-frequency figure is generated, Eugene Wigner-Willie time frequency distribution map of signal is obtained as Fourier transformation to the instantaneous correlation function of signal；(3) time frequency distribution map is pre-processed, generates training sample set and test sample collection；(4) it constructs multiple dimensioned convolutional neural networks model and is trained；(5) test set is tested using trained network model, calculates accuracy, obtain recognition accuracy, assess network performance.The present invention has universality strong, does not need manual features extraction and a large amount of priori knowledges, complexity is low, the advantage that classification results are accurate, stable, can be used in Modulation recognition identification technology field.
Description
Technical field
The invention belongs to signal processing technology fields, further relate to a kind of nothing based on multiple dimensioned convolutional neural networks
Line electric signal Automatic Modulation Recognition method.Present invention may apply to complicated electromagnetic environments, realize the automatic of radio signal
Feature extraction and modulation system classification, to keep the classification of radio signal modulation system more flexible, efficiently.
Background technique
Radio signal Modulation Identification is fought in military electronic, plays important angle in hostile scouting and signal capture analysis
Color, in the case where Given information extremely lacks, first procedure of the Modulation Identification of signal as signal processing flow, to letter
The final identification of breath plays decisive role.Due to the scarcity of prior information, all the time both at home and abroad major scientific research institution and
Colleges and universities are made that a large amount of work in Modulation Identification field.It is identified currently based on the digital signal modulation mode of conventional sorting methods
Satisfactory discrimination can be reached on given test signal.But with the fast development of science and technology, the complexity of electromagnetic environment
The shortcomings that degree improves, and signal kinds and interference increase, these conventional methods is also more prominent.Conventional method needs a large amount of priori
Knowledge and complicated manual features are extracted, and can only identify limited several signals, method robustness is not high and in complex communication ring
It is disturbed and is affected under border, while model is complex.We use multiple dimensioned convolutional neural networks, realize in complicated electromagnetism
Classify under environment to the Automatic Feature Extraction of radio signal and modulation system.
A kind of patent document " CPFSK Modulation Identification method " (application number of University of Electronic Science and Technology in its application
201510847573.2) a kind of Continuous phase frequency shift keying Modulation Identification method is disclosed in, this method passes through following steps: filling
Point using Continuous phase frequency shift keying signal feature, i.e., in each symbol its instantaneous phase be it is linearly increasing or reduce this
One feature, by Continuous phase frequency shift keying signal modeling and extracting signal transient phase and making in conjunction with the means of linear fit
New feature extracting method has better noise resisting ability, by emulation experiment it can be seen that the arithmetic result is intuitive, performance
Well, while possessing lower computational complexity.Shortcoming existing for this method is: although this method proposes a kind of communication
Signal modulate method, but can be only used to identification Continuous phase frequency shift keying signal, and carrying out signal characteristic
A large amount of priori knowledge is needed when extraction.
A kind of patent document " robust communication signal modulate method " (application number of Harbin Institute of Technology in its application
201410680905.8) in disclose a kind of robust communication signal modulate method.This method passes through following steps: 1, to logical
Letter sample of signal carries out Wigner (Wigner-Ville) transformation and obtains time-frequency distributions, extracts second order solid autocorrelation characteristic, builds
Vertical second order solid autocorrelation haracter collection, then carries out selecting the feature set to form robust, later to second order solid autocorrelation characteristic
Training establishes one-class support vector machine group and calculates the output function of one-class support vector machine group.2, signal of communication to be identified is calculated
The probability that sample belongs to the various modulation systems for including in signal of communication sample chooses the modulation class of maximum probability as final
Modulation Identification result.Although this method proposes a kind of robust communication signal modulate method, but this method is still deposited
Shortcoming be: model is complicated, carries out being highly dependent on manual features extraction when signal characteristic abstraction.
Summary of the invention
The present invention in view of the above shortcomings of the prior art, proposes a kind of aerogram based on multiple dimensioned convolutional neural networks
Number Automatic Modulation Recognition method.
Realizing the concrete thought of the object of the invention is, carries out radio signal modulation using multiple dimensioned convolutional neural networks and knows
Not.The algorithm can reach higher discrimination in signal identification, while can reduce conventional modulated recognition methods again to artificial
The high dependency of feature extraction and priori knowledge, can identify the radio signal of multiple types modulation system, and simplify
Identification step.To keep radio signal Modulation Identification more flexible, efficient.
Realize the object of the invention specific steps include the following:
(1) treated radio modulation signal is generated:
By each of the total 220000 radio modulation signals of ten one kind signal by rayleigh fading channel, then fold
Adding signal-to-noise ratio is+5 decibels of white Gaussian noise, obtains 220000 radio modulation signals；
(2) two-dimentional time-frequency figure is generated:
(2a) utilizes Eugene Wigner-Willie time-frequency distributions formula, asks each in 220000 radio modulation signals respectively
The Eugene Wigner of a radio modulation signal-Willie time-frequency distributions；
(2b) draws Eugene Wigner-Willie time-frequency distributions contour map, obtains a kind of ten total 220000 two-dimentional time-frequencies
Figure；
(3) training sample set and test sample collection are generated:
(3a) carries out normalizing to each Zhang Erwei time-frequency figure in a kind of ten two-dimentional time-frequency figures respectively according to normalization formula
Change processing, is combined into image pattern set for the two-dimentional time-frequency figure after all normalizeds；
(3b) randomly selects 80% sample from each two-dimentional time-frequency figure of image pattern set respectively, is combined into instruction
Practice sample set, remaining 20% group is combined into test sample collection；
(4) multiple dimensioned convolutional neural networks model is constructed:
(4a) sets the parameter and maximum number of iterations of multiple dimensioned convolutional neural networks, and maximum number of iterations is set as 100000
Step；
(4b) building is for carrying out 12 layers of convolutional neural networks model of Automatic Feature Extraction to signal；
Two multiple dimensioned convolution for extracting signal Analysis On Multi-scale Features are added in 12 layers of convolutional neural networks model in (4c)
Layer, obtains 14 layers of multiple dimensioned convolutional neural networks model；
Loss function, optimization algorithm, the classifier of multiple dimensioned convolutional neural networks model is arranged in (4d)；
(5) the multiple dimensioned convolutional neural networks model of training:
(5a) concentrates the training sample set of all sample permutations sequences by training sample is upset, and is input to multiple dimensioned convolution mind
Through in network model；
The multiple dimensioned convolutional neural networks model of (5b) training, when the number of iterations for reaching multiple dimensioned convolutional neural networks setting
When, the training process of convolutional neural networks is completed, trained multiple dimensioned convolutional neural networks model is obtained；
(6) recognition accuracy is obtained:
Test sample collection is input in trained multiple dimensioned convolutional neural networks model by (6a), obtains recognition result；
(6b) compares the true classification of recognition result and test set, counts recognition correct rate.
Compared with the prior art, the present invention has the following advantages:
First, it is automatic special for being carried out to radio signal since present invention uses 12 layers of convolutional neural networks models
Sign is extracted, and the shortcomings that prior art needs a large amount of priori knowledges when carrying out radio signal characteristics extraction is overcome.Make this hair
Multiple dimensioned convolutional neural networks model can automatically process the Modulation Mode Recognition of multiple types signal in bright, enhance multiple dimensioned
The universality of convolutional neural networks model.
Second, since two multiple dimensioned convolutional layers are added in the present invention in 12 layers of convolutional neural networks model, obtain 14
The multiple dimensioned convolutional neural networks model of layer increases feature for carrying out the feature extraction of various ways to radio signal
Diversity overcomes the prior art and is highly dependent on the shortcomings that manual features are extracted when carrying out signal analysis, improves simultaneously
Multiple dimensioned convolutional neural networks model simplifies identification step for the accuracy of identification of radio signal modulation system.
Detailed description of the invention
Fig. 1 is flow chart of the invention；
Fig. 2 is a kind of two-dimentional time-frequency figure of the present invention generated ten；
Fig. 3 is analogous diagram of the invention.
Specific embodiment
Invention is described further with reference to the accompanying drawing.
Referring to attached drawing 1, specific steps of the invention are further described.
Step 1, treated radio modulation signal is generated.
By each of the total 220000 radio modulation signals of ten one kind signal by rayleigh fading channel, then fold
Adding signal-to-noise ratio is+5 decibels of white Gaussian noise, obtains 220000 radio modulation signals.
A kind of type of radio modulation signal of described ten is respectively as follows: amplitude modulation double side band signal AMDSB, AM single-side-band
Signal AMSSB, binary phase shift keying modulated signal BPSK, quaternary PSK modulated signal QPSK, eight phase phase-shift keying tune
Signal EPSK processed, Broadband FM signal WBFM, Continuous phase frequency shift keying signal CPFSK, pulse amplitude modulated signal PAM4, ten
Senary quadrature amplitude modulation signal QAM16,60 quaternary quadrature amplitude modulation signal QAM64, Gaussian frequency shifted key signal
GFSK。
Step 2, two-dimentional time-frequency figure is generated.
The first step asks every in 220000 radio modulation signals using Eugene Wigner-Willie time-frequency distributions formula respectively
The Eugene Wigner of one radio modulation signal-Willie time-frequency distributions.
The Eugene Wigner-Willie time-frequency distributions formula is as follows:
Wherein, Wn(t, Ω) indicates n-th of radio modulation signal xn(t) energy at any time t and angular frequency Ω variation
Time-frequency distributions, Ω indicate n-th of radio modulation signal xn(t) angular frequency,Indicate integration operation,It indicates
N-th of radio modulation signal xn(t) existThe value at moment,Indicate n-th of radio modulation signal xn(t) existThe value at moment, τ indicate lag time, and * indicates conjugate operation, and e is indicated using natural logrithm as the index operation of the truth of a matter, j table
Show imaginary unit's symbol.
Second step draws Eugene Wigner-Willie time-frequency distributions contour map, when obtaining a kind of ten total 220000 two dimensions
Frequency is schemed, and a two-dimentional time-frequency figure is taken out from each in a kind of ten two-dimentional time-frequency figures respectively, when taking out 11 two dimensions altogether
Frequency is schemed, and 11 two-dimentional time-frequency figures are as shown in Figure 2.
Step 3, training sample set and test sample collection are generated.
The first step respectively carries out each Zhang Erwei time-frequency figure in a kind of ten two-dimentional time-frequency figures according to normalization formula
Two-dimentional time-frequency figure after all normalizeds is combined into image pattern set by normalized.
The normalization formula is as follows:
Wherein, YmIndicate m two-dimentional time-frequency figure XmImage pattern after normalized, μ indicate a kind of cumulative ten two dimensions
The average value sought after time-frequency figure, σ indicate a kind of ten standard deviations of two-dimentional time-frequency figure.
Second step randomly selects 80% sample, combination from each two-dimentional time-frequency figure of image pattern set respectively
At training sample set, remaining 20% group is combined into test sample collection.
Step 4, multiple dimensioned convolutional neural networks model is constructed.
The first step, sets the parameter and maximum number of iterations of multiple dimensioned convolutional neural networks, and maximum number of iterations is set as
100000 steps.
The parameter setting of the multiple dimensioned convolutional neural networks is as follows: setting 0.001 for learning rate, batch processing size
16 are set as, the image upper limit is read every time and is set as 1000.
Second step constructs 12 layers of convolutional neural networks model for carrying out Automatic Feature Extraction to signal.
The structure setting of 12 layers of convolutional neural networks are as follows: input layer → 1 → pond of convolutional layer, 1 → convolutional layer of layer 2
→ pond 2 → convolutional layer of layer, 3 → pond layer 3 → full articulamentum 1 → full articulamentum 2 → complete 3 → classifier of articulamentum layer → output
Layer.
Wherein the parameter setting of each layer is as follows:
128 neural units are set by input layer.
16 convolution kernels, the window that each convolution kernel is 3 × 3 are set by convolutional layer 1.
Maximum pond is set by pond layer 1, pond layer 2 and pond layer 3 respectively.
16 convolution kernels, the window that each convolution kernel is 3 × 3 is arranged in convolutional layer 2 and convolutional layer 3 respectively.
128 full connection neurons are set by full articulamentum 1 and full articulamentum 2 respectively.
5 full connection neurons are set by full articulamentum 3.
More classification function Softmax are set by classifier layer.
5 output nerve units are arranged in output layer.
Third step, in 12 layers of convolutional neural networks model, two that extraction signal Analysis On Multi-scale Features are added are multiple dimensioned
Convolutional layer obtains 14 layers of multiple dimensioned convolutional neural networks model.
The structure setting of 14 layers of multiple dimensioned convolutional neural networks is input layer → 1 → pond of convolutional layer layer 1 → volume
2 → pond of lamination, 2 → convolutional layer of layer, 3 → pond layer 3 → multiple dimensioned 1 → concatenation of convolutional layer 1 → multiple dimensioned convolutional layer 2 → spelling
Connect operation 2 → full articulamentum 1 → full articulamentum 2 → complete 3 → classifier of articulamentum layer → output layer.14 layers of convolutional neural networks ginseng
In addition to multiple dimensioned convolutional layer and splicing layer, the parameter setting of other each layers and the parameter of 12 layers of convolutional neural networks are set for several settings
Set it is identical, wherein multiple dimensioned convolutional layer and splicing layer parameter setting it is as follows:
Three parallel branches are arranged in multiple dimensioned convolutional layer 1；32 are set by the branch 1 in multiple dimensioned convolutional layer 1
Convolution kernel, the window that each convolution kernel is 1 × 1；Three convolutional layers are set by the branch 2 in multiple dimensioned convolutional layer 1, first
Convolutional layer is set as 32 convolution kernels, the window that each convolution kernel is 1 × 1, and second convolutional layer is set as 24 convolution kernels, often
The window that a convolution kernel is 1 × 1, third convolutional layer are set as 32 convolution kernels, the window that each convolution kernel is 5 × 5；It will be more
Branch 3 in scale convolutional layer 1 is set as three convolutional layers, and first convolutional layer is set as 32 convolution kernels, each convolution kernel
For 1 × 1 window, second convolutional layer is set as 48 convolution kernels, the window that each convolution kernel is 3 × 3, third convolutional layer
It is set as 48 convolution kernels, the window that each convolution kernel is 3 × 3.
Set matrix splicing function for concatenation 1, to the output results of three branches in multiple dimensioned layer 1 into
Row splicing.
Three parallel branches are arranged in multiple dimensioned convolutional layer 2；32 are set by the branch 1 in multiple dimensioned convolutional layer 2
Convolution kernel, the window that each convolution kernel is 1 × 1；Three convolutional layers are set by the branch 1 in multiple dimensioned convolutional layer 2, first
Convolutional layer is set as 32 convolution kernels, the window that each convolution kernel is 1 × 1, and second convolutional layer is set as 16 convolution kernels, often
The window that a convolution kernel is 3 × 3, third convolutional layer are set as 16 convolution kernels, the window that each convolution kernel is 3 × 3；It will be more
Branch 3 in scale convolutional layer 2 is set as maximum pond function.
Set matrix splicing function for concatenation 2, to the output results of three branches in multiple dimensioned layer 2 into
Row splicing.
Loss function, optimization algorithm, the classifier of multiple dimensioned convolutional neural networks model is arranged in 4th step.
The setting of the loss function, optimization algorithm and activation primitive are as follows: set cross entropy for loss function, optimize
Algorithms selection Back Propagation Algorithm sets activation primitive to correct linear unit activating function.
Step 5, the multiple dimensioned convolutional neural networks model of training.
The first step will be upset training sample and be concentrated the training sample set of all sample permutations sequences, is input to multiple dimensioned volume
In product neural network model.
Second step, the multiple dimensioned convolutional neural networks model of training, when the iteration for reaching multiple dimensioned convolutional neural networks setting
When number, the training process of convolutional neural networks is completed, obtains trained multiple dimensioned convolutional neural networks model.
Step 6, recognition accuracy is obtained.
Test sample collection is input in trained multiple dimensioned convolutional neural networks model by the first step, obtains identification knot
Fruit.
Second step compares the true classification of recognition result and test set, counts recognition correct rate.
Effect of the invention can be further illustrated by following emulation:
1. simulated conditions:
Emulation experiment of the invention is in Intel (R) I5-6600K CPU 3.5GHz, GTX1070, Ubuntu16.04LTS
Under system, on TensorFlow1.0.1 operation platform, completes the present invention and radio signal generates and multiple dimensioned convolutional Neural
The emulation experiment of network.
2. emulation experiment content:
By 11 time-frequency distributions images used in emulation experiment of the invention as shown in Fig. 2, the time-frequency distributions image can be with
It is divided into AM single-side-band signal shown in amplitude modulation double side band signal time frequency distribution map, Fig. 2 (b) shown in Fig. 2 (a) by modulation system
Four Xiang Yixiang shown in binary phase shift keying modulated signal time frequency distribution map, Fig. 2 (d) shown in time frequency distribution map, Fig. 2 (c)
Shown in eight phase shift key modulation signal time frequency distribution maps, Fig. 2 (f) shown in keying modulated signal time frequency distribution map, Fig. 2 (e)
Broadband FM signal time frequency distribution map, Continuous phase frequency shift keying signal time frequency distribution map, Fig. 2 (h) institute shown in Fig. 2 (g)
Hexadecimal quadrature amplitude-modulated signal time-frequency distributions shown in the pulse amplitude modulated signal time frequency distribution map shown, Fig. 2 (i)
60 quaternary quadrature amplitude modulation signal time frequency distribution maps shown in figure, Fig. 2 (j), GFSK Gaussian Frequency Shift Keying shown in Fig. 2 (k)
Signal time frequency distribution map.
3. the simulation experiment result:
The simulation experiment result of the invention is as shown in Figure 3.Horizontal axis in Fig. 3 represents the number of iterations, and the longitudinal axis is corresponding to change every time
The loss function value in generation.During to multiple dimensioned convolutional neural networks model training, the loss of each training result is counted
The training effect of functional value, the smaller representative model of loss function value is better.As seen from Figure 3, as the increase of the number of iterations is lost
Functional value successively decreases and finally converges to stabilization, illustrates that the training effect of this emulation experiment is improved with increasing for frequency of training.
Test sample is inputted into trained multiple dimensioned convolutional neural networks model, the identification for obtaining this emulation experiment is accurate
Rate is 95%.
It can be illustrated by above emulation experiment, for the Modulation Identification of radio signal, the present invention can complete difference
The Modulation Identification task of classification, method are effective and feasible.
Claims (8)
1. a kind of radio signal Modulation Identification method based on multiple dimensioned convolutional neural networks, it is characterised in that: including as follows
Step:
(1) treated radio modulation signal is generated:
By each of the total 220000 radio modulation signals of ten one kind signal by rayleigh fading channel, then it is superimposed letter
It makes an uproar than the white Gaussian noise for+5 decibels, obtains 220000 radio modulation signals；
(2) two-dimentional time-frequency figure is generated:
(2a) utilizes Eugene Wigner-Willie time-frequency distributions formula, seeks each of 220000 radio modulation signals nothing respectively
The Eugene Wigner of line electrical modulation signal-Willie time-frequency distributions；
(2b) draws Eugene Wigner-Willie time-frequency distributions contour map, obtains a kind of ten total 220000 two-dimentional time-frequency figures；
(3) training sample set and test sample collection are generated:
Place is normalized to each Zhang Erwei time-frequency figure in a kind of ten two-dimentional time-frequency figures respectively according to normalization formula in (3a)
Reason, is combined into image pattern set for the two-dimentional time-frequency figure after all normalizeds；
(3b) randomly selects 80% sample from each two-dimentional time-frequency figure of image pattern set respectively, is combined into trained sample
This collection, remaining 20% group is combined into test sample collection；
(4) multiple dimensioned convolutional neural networks model is constructed:
(4a) sets the parameter and maximum number of iterations of multiple dimensioned convolutional neural networks, and maximum number of iterations is set as 100000 steps；
(4b) building is for carrying out 12 layers of convolutional neural networks model of Automatic Feature Extraction to signal；
Two multiple dimensioned convolutional layers for extracting signal Analysis On Multi-scale Features are added in 12 layers of convolutional neural networks model in (4c),
Obtain 14 layers of multiple dimensioned convolutional neural networks model；
Loss function, optimization algorithm and the activation primitive of multiple dimensioned convolutional neural networks model is arranged in (4d)；
(5) the multiple dimensioned convolutional neural networks model of training:
(5a) concentrates the training sample set of all sample permutations sequences by training sample is upset, and is input to multiple dimensioned convolutional Neural net
In network model；
The multiple dimensioned convolutional neural networks model of (5b) training, when reaching the number of iterations of multiple dimensioned convolutional neural networks setting,
The training process for completing convolutional neural networks, obtains trained multiple dimensioned convolutional neural networks model；
(6) recognition accuracy is obtained:
Test sample collection is input in trained multiple dimensioned convolutional neural networks model by (6a), obtains recognition result；
(6b) compares the true classification of recognition result and test set, counts recognition correct rate.
2. the radio signal Modulation Identification method according to claim 1 based on multiple dimensioned convolutional neural networks, special
Sign is, ten a kind of types of radio modulation signal described in step (1) are respectively as follows: amplitude modulation double side band signal AMDSB, list
Sideband amplitude-modulated signal AMSSB, binary phase shift keying modulated signal BPSK, quaternary PSK modulated signal QPSK, eight phase shifts
Phase keying modulated signal EPSK, Broadband FM signal WBFM, Continuous phase frequency shift keying signal CPFSK, pulse amplitude modulation letter
Number PAM4, hexadecimal quadrature amplitude-modulated signal QAM16,60 quaternary quadrature amplitude modulation signal QAM64, Gaussian frequency shift
Keying signal GFSK.
3. the radio signal Modulation Identification method according to claim 1 based on multiple dimensioned convolutional neural networks, special
Sign is that Eugene Wigner described in step (2)-Willie time-frequency distributions formula is as follows:
Wherein, Wn(t, Ω) indicates n-th of radio modulation signal xn(t) the energy time-frequency that t and angular frequency Ω changes at any time
Distribution, Ω indicate n-th of radio modulation signal xn(t) angular frequency,Indicate integration operation,Indicate n-th
A radio modulation signal xn(t) existThe value at moment,Indicate n-th of radio modulation signal xn(t) existThe value at moment, τ indicate lag time, and * indicates conjugate operation, and e is indicated using natural logrithm as the index operation of the truth of a matter, j table
Show imaginary unit's symbol.
4. the radio signal Modulation Identification method according to claim 1 based on multiple dimensioned convolutional neural networks, special
Sign is that normalization formula described in step (3a) is as follows:
Wherein, YmIndicate m two-dimentional time-frequency figure XmImage pattern after normalized, μ indicate a kind of cumulative ten two-dimentional time-frequencies
The average value sought after figure, σ indicate a kind of ten standard deviations of two-dimentional time-frequency figure.
5. the radio signal Modulation Identification method according to claim 1 based on multiple dimensioned convolutional neural networks, special
Sign is that the parameter setting of multiple dimensioned convolutional neural networks described in step (4a) is as follows: 0.001 is set by learning rate,
Batch processing is dimensioned to 16, reads the image upper limit every time and is set as 1000.
6. the radio signal Modulation Identification method according to claim 1 based on multiple dimensioned convolutional neural networks, special
Sign is, the structure setting of 12 layers of convolutional neural networks described in step (4b) are as follows: input layer → 1 → pond of convolutional layer layer 1
2 → pond of → convolutional layer, 2 → convolutional layer of layer, 3 → pond layer 3 → full articulamentum 1 → full articulamentum 2 → 3 → classifier of full articulamentum
Layer → output layer；Wherein the parameter setting of each layer is as follows:
128 neural units are set by input layer；
16 convolution kernels, the window that each convolution kernel is 3 × 3 are set by convolutional layer 1；
Maximum pond is set by pond layer 1, pond layer 2 and pond layer 3 respectively；
16 convolution kernels, the window that each convolution kernel is 3 × 3 is arranged in convolutional layer 2 and convolutional layer 3 respectively；
128 full connection neurons are set by full articulamentum 1 and full articulamentum 2 respectively；
5 full connection neurons are set by full articulamentum 3；
More classification function Softmax are set by classifier layer；5 output nerve units are arranged in output layer.
7. the radio signal Modulation Identification method according to claim 1 based on multiple dimensioned convolutional neural networks, special
Sign is that the structure setting of 14 layers of convolutional neural networks described in step (4c) is as follows:
Input layer → 1 → pond of convolutional layer, 1 → convolutional layer of layer, 2 → pond, 2 → convolutional layer of layer, 3 → pond layer 3 → multiple dimensioned convolution
Layer 1 → concatenation 1 → 2 → concatenation of multiple dimensioned convolutional layer 2 → full articulamentum 1 → full articulamentum 2 → full articulamentum 3 → point
Class device layer → output layer；
The setting of 14 layers of convolutional neural networks parameter in addition to multiple dimensioned convolutional layer and concatenation, the parameter setting of other each layers with
The parameter setting of 12 layers of convolutional neural networks is identical, wherein the parameter setting of multiple dimensioned convolutional layer and concatenation is as follows:
Three parallel branches are arranged in multiple dimensioned convolutional layer 1；32 convolution are set by the branch 1 in multiple dimensioned convolutional layer 1
Core, the window that each convolution kernel is 1 × 1；Three convolutional layers, first convolution are set by the branch 2 in multiple dimensioned convolutional layer 1
Layer is set as 32 convolution kernels, the window that each convolution kernel is 1 × 1, and second convolutional layer is set as 24 convolution kernels, Mei Gejuan
The window that product core is 1 × 1, third convolutional layer are set as 32 convolution kernels, the window that each convolution kernel is 5 × 5；It will be multiple dimensioned
Branch 3 in convolutional layer 1 is set as three convolutional layers, and first convolutional layer is set as 32 convolution kernels, and each convolution kernel is 1 ×
1 window, second convolutional layer are set as 48 convolution kernels, the window that each convolution kernel is 3 × 3, the setting of third convolutional layer
For 48 convolution kernels, the window that each convolution kernel is 3 × 3；
Set matrix splicing function for concatenation 1, to the output results of three branches in multiple dimensioned convolutional layer 1 into
Row splicing；
Three parallel branches are arranged in multiple dimensioned convolutional layer 2；32 convolution are set by the branch 1 in multiple dimensioned convolutional layer 2
Core, the window that each convolution kernel is 1 × 1；Three convolutional layers, first convolution are set by the branch 1 in multiple dimensioned convolutional layer 2
Layer is set as 32 convolution kernels, the window that each convolution kernel is 1 × 1, and second convolutional layer is set as 16 convolution kernels, Mei Gejuan
The window that product core is 3 × 3, third convolutional layer are set as 16 convolution kernels, the window that each convolution kernel is 3 × 3；It will be multiple dimensioned
Branch 3 in convolutional layer 2 is set as maximum pond function；
Set matrix splicing function for concatenation 2, to the output results of three branches in multiple dimensioned convolutional layer 2 into
Row splicing.
8. the radio signal Modulation Identification method according to claim 1 based on multiple dimensioned convolutional neural networks, special
Sign is, loss function described in step (4d), the setting of optimization algorithm and activation primitive are as follows: be set as handing over by loss function
Entropy is pitched, optimization algorithm Select Error sets activation primitive to correct linear unit activating function against propagation algorithm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711144077.6A CN107979554B (en) | 2017-11-17 | 2017-11-17 | Radio signal Modulation Identification method based on multiple dimensioned convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711144077.6A CN107979554B (en) | 2017-11-17 | 2017-11-17 | Radio signal Modulation Identification method based on multiple dimensioned convolutional neural networks |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107979554A CN107979554A (en) | 2018-05-01 |
| CN107979554B true CN107979554B (en) | 2019-10-08 |
Family
ID=62010285
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711144077.6A Active CN107979554B (en) | 2017-11-17 | 2017-11-17 | Radio signal Modulation Identification method based on multiple dimensioned convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107979554B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11826129B2 (en) * | 2019-10-07 | 2023-11-28 | Owlet Baby Care, Inc. | Heart rate prediction from a photoplethysmogram |
Families Citing this family (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108650202B (en) * | 2018-05-11 | 2021-02-26 | 大唐联诚信息系统技术有限公司 | Signal modulation mode identification method and device |
| CN109120563B (en) * | 2018-08-06 | 2020-12-29 | 电子科技大学 | Modulation recognition method based on neural network integration |
| CN109271926B (en) * | 2018-09-14 | 2021-09-10 | 西安电子科技大学 | Intelligent radiation source identification method based on GRU deep convolutional network |
| CN109214356A (en) * | 2018-09-29 | 2019-01-15 | 南京东振测控技术有限公司 | A kind of fan transmission system intelligent fault diagnosis method based on DCNN model |
| CN109299697A (en) * | 2018-09-30 | 2019-02-01 | 泰山学院 | Deep neural network system and method based on underwater sound communication Modulation Mode Recognition |
| CN109347601B (en) * | 2018-10-12 | 2021-03-16 | 哈尔滨工业大学 | Convolutional neural network-based decoding method of anti-tone-interference LDPC code |
| CN109787927A (en) * | 2019-01-03 | 2019-05-21 | 荆门博谦信息科技有限公司 | Modulation Identification method and apparatus based on deep learning |
| CN109787929A (en) * | 2019-02-20 | 2019-05-21 | 深圳市宝链人工智能科技有限公司 | Signal modulate method, electronic device and computer readable storage medium |
| CN110417694A (en) * | 2019-06-07 | 2019-11-05 | 吉林大学 | A kind of modulation mode of communication signal recognition methods |
| CN111404852B (en) * | 2020-03-03 | 2021-04-13 | 西安电子科技大学 | Modulation mode identification method based on amplitude and spectral amplitude characteristics |
| CN111401263B (en) * | 2020-03-19 | 2022-07-08 | 电子科技大学 | Expert knowledge fused optimal effect combined modulation identification method |
| CN111709329B (en) * | 2020-05-31 | 2023-04-28 | 中国人民解放军63892部队 | Unmanned aerial vehicle measurement and control signal high-speed recognition method based on deep learning |
| CN111814953B (en) * | 2020-06-16 | 2024-02-13 | 上海瀚讯信息技术股份有限公司 | Positioning method of deep convolution neural network model based on channel pruning |
| CN112580598B (en) * | 2020-12-30 | 2024-02-13 | 浙江工业大学 | Radio signal classification method based on multichannel Diffpool |
| CN112784690B (en) * | 2020-12-31 | 2022-12-27 | 西安电子科技大学 | Broadband signal parameter estimation method based on deep learning |
| CN112818876B (en) * | 2021-02-04 | 2022-09-20 | 成都理工大学 | Electromagnetic signal extraction and processing method based on deep convolutional neural network |
| CN112887239B (en) * | 2021-02-15 | 2022-04-26 | 青岛科技大学 | Method for rapidly and accurately identifying underwater sound signal modulation mode based on deep hybrid neural network |
| CN113095137B (en) * | 2021-03-10 | 2023-12-29 | 南京大学 | Signal characteristic recognition device and method based on machine learning |
| CN113542180A (en) * | 2021-06-30 | 2021-10-22 | 北京频谱视觉科技有限公司 | Frequency domain identification method of radio signal |
| CN113723556B (en) * | 2021-09-08 | 2022-05-31 | 中国人民解放军国防科技大学 | Modulation mode identification method based on entropy weighting-multi-mode domain antagonistic neural network |
| CN113783815B (en) * | 2021-09-10 | 2022-12-06 | 苏州大学 | User-defined convolutional neural network modulation identification method and system for wireless communication |
| CN114429156B (en) * | 2022-01-21 | 2023-06-20 | 西安电子科技大学 | Radar interference multi-domain characteristic countermeasure learning and detection recognition method |
| CN115720184B (en) * | 2022-10-08 | 2024-04-19 | 西安电子科技大学 | Small sample signal modulation type identification method based on characteristic distribution |
| CN117857270B (en) * | 2024-03-07 | 2024-05-14 | 四川广播电视监测中心 | Radio communication signal modulation identification method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105956532A (en) * | 2016-04-25 | 2016-09-21 | 大连理工大学 | Traffic scene classification method based on multi-scale convolution neural network |
| CN106599941A (en) * | 2016-12-12 | 2017-04-26 | 西安电子科技大学 | Method for identifying handwritten numbers based on convolutional neural network and support vector machine |
| CN107122738A (en) * | 2017-04-26 | 2017-09-01 | 成都蓝色起源科技有限公司 | Automatic Communication Signals Recognition based on deep learning model and its realize system |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9779492B1 (en) * | 2016-03-15 | 2017-10-03 | International Business Machines Corporation | Retinal image quality assessment, error identification and automatic quality correction |
-
2017
- 2017-11-17 CN CN201711144077.6A patent/CN107979554B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105956532A (en) * | 2016-04-25 | 2016-09-21 | 大连理工大学 | Traffic scene classification method based on multi-scale convolution neural network |
| CN106599941A (en) * | 2016-12-12 | 2017-04-26 | 西安电子科技大学 | Method for identifying handwritten numbers based on convolutional neural network and support vector machine |
| CN107122738A (en) * | 2017-04-26 | 2017-09-01 | 成都蓝色起源科技有限公司 | Automatic Communication Signals Recognition based on deep learning model and its realize system |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11826129B2 (en) * | 2019-10-07 | 2023-11-28 | Owlet Baby Care, Inc. | Heart rate prediction from a photoplethysmogram |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107979554A (en) | 2018-05-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107979554B (en) | Radio signal Modulation Identification method based on multiple dimensioned convolutional neural networks | |
| CN108234370B (en) | Communication signal modulation mode identification method based on convolutional neural network | |
| CN110855591B (en) | QAM and PSK signal intra-class modulation classification method based on convolutional neural network structure | |
| CN108282263B (en) | Coded modulation joint recognition methods based on one-dimensional depth residual error light weight network | |
| CN110070101B (en) | Plant species identification method and device, storage medium and computer equipment | |
| CN108038471A (en) | A kind of underwater sound communication signal type Identification method based on depth learning technology | |
| CN104166859B (en) | Based on SSAE and FSALS SVM Classification of Polarimetric SAR Image | |
| CN110163282A (en) | Modulation Mode Recognition method based on deep learning | |
| CN109299697A (en) | Deep neural network system and method based on underwater sound communication Modulation Mode Recognition | |
| CN108243129A (en) | Radio signal characteristics extracting method based on attention depth network | |
| CN106446942A (en) | Crop disease identification method based on incremental learning | |
| CN106203625A (en) | A kind of deep-neural-network training method based on multiple pre-training | |
| CN112069883B (en) | Deep learning signal classification method integrating one-dimensional two-dimensional convolutional neural network | |
| CN110175560A (en) | A kind of radar signal intra-pulse modulation recognition methods | |
| CN112702294A (en) | Modulation recognition method for multi-level feature extraction based on deep learning | |
| CN106600595A (en) | Human body characteristic dimension automatic measuring method based on artificial intelligence algorithm | |
| CN110532932A (en) | A kind of multi -components radar emitter signal intra-pulse modulation mode recognition methods | |
| CN110309854A (en) | A kind of signal modulation mode recognition methods and device | |
| CN108230233A (en) | Data enhancing, treating method and apparatus, electronic equipment and computer storage media | |
| CN106611423B (en) | SAR image segmentation method based on ridge ripple filter and deconvolution structural model | |
| CN105095836B (en) | A kind of skin-texture detection method and device based on Gabor characteristic | |
| CN110245711A (en) | The SAR target identification method for generating network is rotated based on angle | |
| CN113378644B (en) | Method for defending signal modulation type recognition attack based on generation type countermeasure network | |
| CN106778910A (en) | Deep learning system and method based on local training | |
| CN110048978A (en) | A kind of signal modulate method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |